Perceiving the Focal Point of a Painting with AI: Case Studies on Works
of Luc Tuymans
Luc Steels
a
and Bj
¨
orn Wahle
Institut de Biologia Evolutiva, Universitat Pompeu Fabra and CSIC, Barcelona, Spain
Keywords:
Digital Humanities, Contemporary Painting, AI, Computer Vision, Iconography.
Abstract:
We report the first steps in investigating how we can use AI to study contemporary painting practices and
viewer experiences, focusing in particular on the work of Luc Tuymans. We review first various possible
aspects of painting that could be studied and point to some relevant AI techniques to do so. Then we zoom in
on one specific topic: How is a viewer guided to the focal point of the painting. This is not purely a matter
of visual perception but also of interpretation and meaning making. Painters deliberately create focal points
based on sophisticated knowledge of human perception and interpretation. Inspired by their insights and
practices we can use AI research to unpack the process and thus provide a more insightful characterization of
how paintings are perceived and made, compared to statistically derived embeddings. We argue that profound
challenges must still be overcome before AI systems handle the identification of focal points, let alone arrive
at the rich interpretations human viewers construct of paintings or other types of art works.
1 INTRODUCTION
The use of AI in digital humanities is increasing
steadily with important and successful applications
for managing and searching in existing collections or
performing historical art research. This work rests on
various AI techniques from pattern recognition, com-
puter vision, and information retrieval, augmented
with semantic web technologies ((Strezoski and Wor-
ring, 2017), (de Boer et al., 2013)) Other interac-
tions between AI and art have focused on capturing
the characteristics of artistic style in order to generate
new art works in the same style (Semmo et al., 2017).
Here we report on a rather different line of work
that is exploring how AI can model the process of per-
ceiving and interpreting the meaning of art works, in-
dividually or in context, i.e. the semantic understand-
ing of art (Garcia and Vogiatzis, 2018) We want to un-
pack the perceptual and interpretive processes that a
viewer goes through, using AI models as a vehicle for
examining these processes and studying their effect.
This has educational applications to help viewers have
richer experiences but also it is of interest to painters
because they can explore the effect of their work in a
novel way and possibly open innovative paths in the
creation of new art works. The AI systems and exper-
a
https://orcid.org/0000-0001-9134-3663
imental outcomes of our approach are a valuable addi-
tion to the traditional data gathered about art and can
be a complementary instrument for heritage preserva-
tion.
The present research is being conducted in inter-
action with the contemporary Flemish painter Luc
Tuymans. The role of a professional artist is cru-
cial for our project. We view him as a highly com-
petent expert in human perception and interpretation.
He uses these insights to create works that maximize
the artistic experience and achieve rich meaning con-
struction for his viewers. There are considerable ad-
vantages in working with a living artist because we
can get much more accurate data about the work, the
context of creation, the source materials of each work,
the texts written by curators or the artist himself, the
exhibition design process, and the intended meanings
(from the viewpoint of the artist - which may differ
from those of viewers). We can also get direct feed-
back about the results of running various AI experi-
ments and learn what the painter finds important and
relevant for the study of artistic practices.
The main source of data for the present research is
the complete collection of paintings of Tuymans (564
works), together with extensive meta-data methodi-
cally archived digitally and compiled in a ‘Catalogue
Raisonn
´
e’ (Meyer-Hermann, 2019), from which we
have extracted high quality digital images for each
Steels, L. and Wahle, B.
Perceiving the Focal Point of a Painting with AI: Case Studies on Works of Luc Tuymans.
DOI: 10.5220/0009163108950901
In Proceedings of the 12th International Conference on Agents and Artificial Intelligence (ICAART 2020) - Volume 2, pages 895-901
ISBN: 978-989-758-395-7; ISSN: 2184-433X
Copyright
c
2022 by SCITEPRESS Science and Technology Publications, Lda. All rights reserved
895
painting. From the digital archive maintained at Stu-
dio Luc Tuymans we have extracted other information
such as in which exhibitions paintings were shown, so
that we could explore the curatorial practices of Tuy-
mans through network analysis. This research is how-
ever not discussed in the present paper.
Our choice to work with Luc Tuymans has already
turned out to be very productive. The mastery and
artistic importance of Tuymans is not in dispute. He
has exhibited in Tate Modern London, Centre for Fine
Arts (BOZAR) Brussels, MOMA New York, Palazzo
Grassi Venice, and many other prestigious venues.
His work has a strong recognition in the art market
through the influential galeries of David Zwirner in
New York and Frank De Maegd (Zeno X) in Antwerp.
It is also very helpful that Tuymans is very articulate
in describing his artistic practice and there is an abun-
dant literature about his art, including interviews, cat-
alogs, art criticisms, and personal writings (Tuymans,
2018).
Even more importantly, the work of Tuymans is
rich at many levels, particularly the conceptual level.
Despite a calm and esthetically pleasing appearance,
his paintings always hide a deeper level of meaning
which challenges us to go beyond pattern recognition
and machine vision in order to integrate meaning and
understanding. Because meaning and understanding
are still open problems for AI, this project is therefore
interesting because it helps to push forward AI be-
yond the current state of the art, which is all too often
focusing on surface characterics of human experience
rather than meaning and understanding.
2 FORM AND MEANING
The famous art historian and semiotician Ervin Panof-
sky identified five levels in the appreciation of art
works, going from form to meaning (Panofsky, 1972).
These same levels can also be distinguished when per-
ceiving and interpreting every-day activities or im-
ages which are not art, and so results of the present
investigation have a wide applicability beyond the art
context.
Each of Panofsky’s levels has already been studied
in depth using AI methods, although as we move from
form to meaning, results are becoming harder to ob-
tain. Moreover there is clearly a tight interaction be-
tween the levels, requiring a bottom-up and top-down
flow of information, which is often not yet adequately
handled by AI architectures.
1. At the bottom level we focus on the formal ap-
pearance of an art work: the colors, lines, and vol-
umes that are perceived by low-level visual processes
and aggregated into coherent segments. These pro-
cesses have been studied intensely in AI (specifically
the fields of pattern recognition and computer vision)
during the past decades, either by designing and im-
plementing feature extractors and pattern detectors
or, more recently, by using some variant of convo-
lutional networks acquiring features and patterns di-
rectly from data (Cetinica et al., 2018). There are
now many libraries of ready-made low-level visual
processing components available and we have already
applied some of them to the paintings in the Tuymans
collection (occasionally after first training the neural
networks involved).
The results we have obtained so far are often quite
unexpected because paintings are not the same as the
kind of pictures with which pattern recognition algo-
rithms are usually trained. Often the original source
image is deliberately distorted or blurred so that basic
low level vision processing, such as edge detection
or shape from shading, is difficult for existing algo-
rithms.
1110
LTP 385
The Book, 2007
Oil on canvas
306 × 212 cm | 120 ½ × 83 ½ inches
Signed and dated on verso, right: “Luc Tuymans 007”
Pinault Collection
PROVENANCE
Zeno X Gallery, Antwerp
SOLO EXHIBITIONS
Les Revenants, Zeno X Storage, Antwerp, 2007.
Wenn der Frühling kommt, Haus der Kunst, Munich, 2008.
La Pelle, Palazzo Grassi, Venice, 2019–20.
SELECTED GROUP EXHIBITIONS
Mapping the Studio: Artisti dalla collezione François Pinault/Artists from
the François Pinault Collection, Punta della Dogana and Palazzo
Grassi, Venice, 2009–11.
MONOGRAPHS AND SOLO EXHIBITION CATALOGUES
Norio Sugawara, Luc Tuymans: Beyond Schwarzheide, ill. 8. Tokyo:
Wako Works of Art, 2007.
Pablo Sigg and Tommy Simoens, ed., Luc Tuymans: Is It Safe?, 126; ill.
72. London: Phaidon, 2010.
Patrizia Dander and Donna Wingate, ed., Luc Tuymans: Wenn der
Frühling kommt, 9, 16; ill. 16. Ed. cat. Haus der Kunst, Munich,
2008. Antwerp: Ludion, 2014.
Frank Demaegd, ed., Luc Tuymans: Zeno X Gallery, 25 Years of
Collaboration, ill. 119, 124, 269. Exh. cat. Antwerp: Zeno X Books,
2016.
SELECTED BOOKS AND GROUP EXHIBITION CATALOGUES
Francesco Bonami and Alison M. Gingeras, ed., Mapping the Studio:
Artisti dalla collezione François Pinault/Artists from the François
Pinault Collection/Artistes de la collection de François Pinault, ill.
272–73. Exh. cat. Venice: Palazzo Grassi; Milan: Mondadori Electa,
2009.
SELECTED LITERATURE
Danny Ilegems, “‘Mijn schilderijen zijn geen schilderijen’: Interview
kunstenaar Luc Tuymans.” Vrij Nederland, May 2007, 70, ill. 70.
Hans Theys, “Van oude spoken en dingen die niet voorbijgaan.” H Art,
May 2007, 3.
Dorine Esser, “‘Ik slaag er niet in iets vrolijks te schilderen.’” Isel, May/
June 2007, ill. 27.
Michele Robecchi, “Luc Tuymans.” Flash Art, October 2007, 130–32.
Yasmine Van Pee, “Unnatural Resources: Luc Tuymans on Fighting the
Literal and Mistrusting Images.” Modern Painters, October 2007,
ill. 75.
Jan Koenot, “De macht van de jezuïeten en de onmacht van beelden:
Terugblik op Luc Tuymans’ serie ‘Les Revenants.’” Streven,
November 2007, 874; ill. 875.
Cornelia Gockel, “Luc Tuymans: Wenn der Frühling kommt.”
Kunstforum International, May–July 2008, 359.
Heinz-Norbert Jocks, “Das Auauchende und das Verschwindende:
Ein Gespräch mit Heinz-Norbert Jocks.” Kunstforum International,
January–February 2019, ill. 194.
NOTES
Part of the group Les Revenants.
Figure 1: Left: ‘The Book’ by Luc Tuymans, oil on canvas
306 X 212 cm, 2017, Pinault Collection. Right: Source
object of this picture, the interior of the Chiesa del Ges
`
u in
Rome. The painting is based on two pages in a book which
contain an image of the church and you see the fold line in
the middle.
Compare for example the Tuymans painting shown in
Fig. 1, left and a photographic picture of the same
scene in Fig. 1 on the right. The painting actually de-
picts a book opened at a picture of the church. The
distortion due to the folding of the pages of the book
and the blurring of lines and surfaces makes low-level
visual processing and image recognition challenging
for humans and even harder for machines. For ex-
ample, standard edge detection and segmentation al-
gorithms are led astray here by the fold mark in the
middle, which is unrelated to the image of the church
and cuts across the whole image.
ICAART 2020 - 12th International Conference on Agents and Artificial Intelligence
896
2. On the second level there is the recognition of the
objects and events that are being depicted in the im-
age: the face of a person, a church interior, the pic-
ture of a clown carrying balloons. Panofsky calls this
level the first stage of meaning, specifically the fac-
tual meaning. For example, the colors, patches of
light and dark, and the shapes in Fig.2(left) quickly
assemble in our perceptual field into the face of a
woman.
325324
LTP 538
Twenty Seventeen, 2017
Oil on canvas
94.7 × 62.7 cm | 37 ¼ × 24 ¾ inches
Signed and dated on verso, top right: “L Tuymans 0017”
Pinault Collection
PROVENANCE
Zeno X Gallery, Antwerp
SOLO EXHIBITIONS
La Pelle, Palazzo Grassi, Venice, 2019–20.
NOTES
The motif of Twenty Seventeen was painted by the artist as a
temporary mural for the exhibition Jiwa, Jakarta Biennale, 2017.
MSI-net
Figure 2: Left: Twenty Seventeen, 2017, oil on canvas 94.7
X 62.7 cm Pinault Collection. It represents the face of a
woman at the moment she learns that she will be put to death
by poisoning. Right: The most salient area according to the
MSI algorithm is overlayed on the painting using red color
- see section 3. of this paper).
We recognize factual meaning based on our abundant
prior visual experiences of the world. This process is
called image recognition in AI research. It is another
area in which there has been considerable progress
lately by training associative neural networks with
millions of labeled segmented images until a network
is obtained that can autonomously segment and label
novel images. There have been several experiments
to apply such image recognition systems to paintings.
But, similar as to low-level visual processing, the re-
sults typically do not quite reach the performance lev-
els compared to what one gets from analyzing every
day photographs or videos with which these networks
have been trained. The miscategorizations of image
recognition algorithms have even become the object
of new art works (Schmitt, 2018).
It is nevertheless very interesting to wonder how
and why algorithmic results deviate from human ex-
pectations. The main reason is that source images
have been blurred, cut out, enlarged or reduced in the
paintings, making them more abstract, timeless, ex-
pressive and therefore iconic.
Surprisingly, the creative interpretation of the
original sources can expose unexpected unconscious
associations. An example from our image recogni-
tion experiments on the Tuymans collection is shown
in Fig. 3. The painting is overlayed with the results of
two state-of-the-art algorithms: the YOLO algorithm
(left) (Redmon et al., 2015) and the Mask R-CNN al-
gorithm (right) (Girshick et al., 2013). The Yolo al-
gorithm labels the segmented person as a dog (with
0.6 certainty) and the Mask R-CNN algorithm labels
the figure as a person (0.6 certainty) but also as a dog
(0.75 certainty). This is at first bizar until we real-
ize Tuymans depicts here a famous Japanese criminal
Issei Sagawa who was a cannibal but managed to es-
cape justice. Although he is not explicitly depicted as
a dog, there are apparently sufficient dog-like features
to push the image recognition algorithms towards this
classification.
YOLO Mask(R-CNN
0.60(dog
0.75(dog
0.60(person
Figure 3: ‘Issei Sagawa’ by Luc Tuymans, 2014. Oil on
canvas 74,3 x 81,9 cm. Tate collection. The application
of the YOLO image recognition algorithm is shown on the
left and of the Mask R-CNN algorithm on the right. Both
algorithms assign the label dog to this human figure.
3. At the next layer, human observers interpret ob-
jects and events in terms of psychological nuances,
such as emotional states of the persons depicted or
the nature of the actions they carry out (aggressive,
friendly). Panofsky calls this the expressional mean-
ing of an art work. For example, a human viewer rec-
ognizes immediately in the painting in Fig. 2 that this
depicts not just the face of a woman, but a woman
who is very surprised or shocked, maybe by some-
thing that she sees or has just heard. She is in distress
and shows anguish.
Detecting expressional meaning using AI is still
more difficult than object recognition. There has been
some work on sentiment analysis of visual images
(You et al., 2015) but this has not at all reached the
depth with which humans are able to grasp expressive
meaning.
4. The fourth level of interpretation requires un-
derstanding who or what is depicted in order to un-
derstand the motivations and intentions of those be-
ing represented, or the situations in which they find
themselvwes. For example, Fig. 1 is an image of an
image in a book of a baroque church interior, namely
the Chiesa del Ges
`
u, the mother church of the Jesuits
in Rome. The woman in Fig. 2 is a character from
Perceiving the Focal Point of a Painting with AI: Case Studies on Works of Luc Tuymans
897
1312
LTP 386
The Valley, 2007
Oil on canvas
106.5 × 109.5 cm | 41 × 43 inches
Signed and dated on verso, top right: “Luc Tuymans 007”
Pinault Collection
PROVENANCE
Zeno X Gallery, Antwerp
SOLO EXHIBITIONS
Les Revenants, Zeno X Storage, Antwerp, 2007.
Wenn der Frühling kommt, Haus der Kunst, Munich, 2008.
La Pelle, Palazzo Grassi, Venice, 2019–20.
MONOGRAPHS AND SOLO EXHIBITION CATALOGUES
Norio Sugawara, Luc Tuymans: Beyond Schwarzheide, ill. iii. Tokyo:
Wako Works of Art, 2007.
Pablo Sigg and Tommy Simoens, ed., Luc Tuymans: Is It Safe?, 69,
127–28; cover, ill. 71. London: Phaidon, 2010.
Patrizia Dander and Donna Wingate, ed., Luc Tuymans: Wenn der
Frühling kommt, 9, 14; ill. 14. Ed. cat. Haus der Kunst, Munich,
2008. Antwerp: Ludion, 2014.
Frank Demaegd, ed., Luc Tuymans: Zeno X Gallery, 25 Years of
Collaboration, ill. 117, 124, 269. Exh. cat. Antwerp: Zeno X Books,
2016.
SELECTED LITERATURE
Hanno Rauterberg, “Schwach gemalt, schwach gedacht.” Die Zeit,
April 24, 2003.
Eric Rinckhout, “Tuymans schildert de macht der jezuïeten.” De
Morgen, April 21, 2007.
Frank Heirman, “Uitverkocht voor opening.” Gazet van Antwerpen,
April 25, 2007.
Jan Van Hove, “De jezuïetenstreken van Luc Tuymans.” De Standaard,
April 25, 2007.
Wim Daneels, “De jezuïetenstreken van Luc Tuymans.” Het Nieuwsblad,
April 26, 2007.
Peter van Dyck, “Luc Tuymans: De duurste Belg.” Gentleman,
April 2007, ill. 52.
Hans Theys, “Van oude spoken en dingen die niet voorbijgaan.”
H Art, May 2007, 3.
Dorine Esser, “‘Ik slaag er niet in iets vrolijks te schilderen.’” Isel,
May/June 2007, 25; ill. 25.
Jeroen Laureyns, “Geschilderde geruchten.” Knack, June 6, 2007.
Jan Koenot, “De macht van de jezuïeten en de onmacht van beelden:
Terugblik op Luc Tuymans’ serie ‘Les Revenants.’” Streven,
November 2007, 870.
“Luc Tuymans on the Failure of Utopias.” Art World, February/March
2008, ill. 16.
Rüdiger Heinze, “Malen auf des Messers Schneide.” Augsburger
Allgemeine, March 1, 2008.
Swantje Karich, “Wieviel Rätsel braucht die Geschichte?” Frankfurter
Allgemeine Zeitung, March 12, 2008.
Thomas Schönberger, “Leerstellen der Monstrosität.” Spex,
March 2008, ill. 126.
Gesine Borcherdt, “‘Wenn der Frühling kommt’ ist der Titel einer
grossen Werkschau im
Münchner Haus der Kunst.” Monopol, April 2008, 115; ill. 115.
Susanna C. Ott, “Geronnene Erinnerung.” Applaus, April 2008.
Morgan Falconer, “Luc Tuymans: Agent Provocateur.” Art World,
April/May 2008, ill. 43.
Hanno Rauterberg, “Was bedeuten diese Bilder?” Die Zeit, May 8,
2008.
Stefanie de Jonge, “De 7 hoofdzonden volgens Luc Tuymans.” Humo,
February 14, 2011, ill. 128.
Elisabeth Vedrenne, “Le monde mental de Luc Tuymans.”
Connaissance des arts, October 2016, ill. 58.
David Castenfors, “Målar Mästare Tuymans.” Artlover, no. 30, 2016,
ill. 5.
NOTES
Part of the group Les Revenants.
Figure 4: Left: ‘The Valley’ by Luc Tuymans, 2007. Oil
on canvas. 106.5 X 109.5 cm. Pinault Collection. Right:
Source image, a still from the Film ’The Valley of the
damned’ directed by Wolf Rilla, 1960. Blurring or lack
of contrast makes low-level image processing and image
recognition very challenging.
a Brazilian dystopian Netflix series called 3 % where
only few people can get to an offshore heaven-like is-
land, where the elite lives, by winning in a game. But
if they lose the game they are killed. This woman just
heard that she lost and will die by poisoning, hence
the expression of shock and dispair. The boy depicted
in Fig. 4 is from a 1960s movie based on the book
’Village of the damned’ by John Wyndham. It is an-
other dystopian story: All the residents of the village
of Midwich become suddenly unconscious for several
hours. Months later, twelve local women and girls
give birth the same day to albino children with phos-
phorescent eyes. Precocious and able to communi-
cate by telepathy they will quickly reveal hostile in-
tentions. (Donnadieu, 2019) The boy shown in the
painting is one of these children.
Panofsky calls this the level of conventional mean-
ing because it rests on knowing conventions in society
and knowing about historical events, well known fig-
ures, and cultural artefacts, like books or films. This
kind of meaning is imposed on the image by calling
on episodic and semantic memory or consulting ex-
ternal knowledge sources such as found on the web.
The AI methods that now come into play are based
on knowledge representation, reasoning, and seman-
tic memory, such as knowledge graphs that store vast
amounts of facts (more than 7 billion for the Google
knowledge graph). No efforts have been made so far
to model interpretation of conventional meanings in
paintings using AI, but it is not excluded to start tack-
ling this issue with current semantic technologies.
5. Finally there is the highest layer which Panof-
sky calls the intrinsic meaning or content of an art
work. Here we address the ultimate motives of the
artist, which could be political, psychological, histor-
ical, or mere story telling.
For example, the paintings shown in Figures
1 and 4 are both coming from an exhibition Les
Revenants about the enduring influence of religious
power, against which the painter wants to rebel, in
particular the influence of the Jesuits. The link to the
Jesuits is straightforward for Fig.1 because it depicts
a Jesuit church. The intrinsic meaning is established
by the distortion and the choice of “yellow, earthy and
greyish white, slightly fuzzy hues, (...) which disrupts
the splendor and magnificence of the place. In so
doing, Luc Tuymans inverts the illusionist character
of religious architecture by blurring the sculpted and
painted representations meant to inspire the faithful
and strengthen their faith. (Donnadieu, 2019).
And even though the boy in Fig. 4 comes from a
totally different context, namely the movie ‘The Val-
ley of the damned’, it still makes a link to the Jesuits.
“The stern, stubborn gaze of the portrayed child, his
strict haircut and clothes signal harsh educational and
social norms or quasi-military upbringing. (Don-
nadieu, 2019). This is indirectly associated with the
Jesuit educational system that wanted to form ‘sol-
diers of Christ’.
In general, the intrinsic meanings conjured up by
Tuymans create feelings of uneasiness and fear by
referring to antagonistic themes, such as the Nazi
regime, racism, child abuse, religious power, crime,
etc., and it is achieved by the selection of source im-
ages, the very constrained use of colors, image dis-
tortions, and close-ups. For example, the face in
Fig. 2 is actually the projection of a face on a doll’s
head, in order to evoke a feeling of alienation and
dystopia. Building AI systems that can handle this
level of meaning is currently totally beyond the state
of the art. Indeed, the more we get to the level of
intrinsic meaning, the more helpless current AI tech-
niques become.
The different processes going from form to mean-
ing through these five levels are not only supported by
the visual aspects of the work which, partially uncon-
sciously, affect the psychological state of the viewer.
But the title of each work and the explanations in the
catalog are also important factors that influence mean-
ing making. The painting shown in Fig.4 is called
‘The Valley’ giving a clue about the source of this
image, namely a movie with the same name. The
painting in Fig.1 is called ‘The book’, thus giving an
important cue that we are looking at an image of a
book, whereas at first one sees a scrambled interior of
a church. The title ‘twentyseventeen’ (Fig.2) refers to
the year 2017 in which the painting (and the movie)
were made. It is a dark year according to Tuymans
with the rise of populism, Brexit, and disinformation
on social media through companies like Cambridge
Analytica, so that fear and dystopia are appropriate
themes to be evoked in that year.
ICAART 2020 - 12th International Conference on Agents and Artificial Intelligence
898
3 THE FOCAL POINT IN
PAINTING
We now zoom in on one of the painterly devices that
shape the perception and interpretation of an art work
and is unique to painting, namely the presence of a
focal point, also known as the entry point or breaking
point. “The focal point of a painting is an area of em-
phasis that demands the most attention and to which
the viewer’s eye is drawn, pulling it into the painting.
(Marder, 2019) The focal point is not accidental. It
is deliberately chosen by the painter in order to help
achieve intended meanings and it partially shapes the
visual experience of the viewer. There can be a few
focal points in a single painting (rarely more than
3). Abstract paintings, which do not convey mean-
ing (such as by Jackson Pollock for example) might
have no focal point at all.
Modeling how AI systems can recognize the fo-
cal point is an interesting, although very challenging,
case study because it involves in principle all levels
of interpretation. Painters use a diverse set of artis-
tic means to guide viewers: Lines, shapes, color dif-
ferences (in hue, value (brightness) and saturation),
textures, space, and composition, as well as title, ex-
hibition context, and narratives found in the catalog.
The first step is to understand the role of natural
perceptual processes, of which one example is the de-
tection of the most salient areas in a painting. We
have therefore concentrated on this process in a first
preliminary study, using state-of-the-art saliency es-
timation algorithms from the computer vision litera-
ture, which contains hundreds of proposals for this
task. Our goal is not to develop new algorithms nor
to train algorithms with new data, but to see how the
best state-of-the-art saliency algorithms and models
identify as salient in the Tuymans paintings.
The algorithms we applied fall in two classes:
Knowledge-free saliency estimation exploits gen-
eral properties of the visual image but does not
take into account statistical features learned from
prior visual experiences nor higher-level semantic
features. It is typical for earlier work in pattern-
recognition. No training is needed to apply these
methods. We have tried two state-of-the-art algo-
rithms on the Tuymans collection:
1. The Spectral Residue Based Method (SRB)
(Hou and Zhang, 2007) which analyzes the log-
spectrum of the image to extract the spectral
residue and computes a saliency map on that
basis.
2. The Fine-Grained Method (Montabone and
Soto, 2010) which is based color constancy
detection and pair callibration, segmentation
based on depth continuity, and visual saliency
based on extracted features. The method is spe-
cialized in recognizing humans.
Knowledge-based Saliency Estimation uses neu-
ral networks that have been trained using super-
vised (deep) learning. We have tried two state-of-
the-art algorithms:
1. POOLNET does salient object detection (not
just region detection) and gives no result when
an object could not be detected (Liu et al.,
2019). It uses a convolutional neural network
with additional components for combining low-
level (close to the visual form) and high-level
(semantic) features with results from other vi-
sual processes such as edge detection. POOL-
NET is trained on a corpus of real world images
of which the salient regions have been anno-
tated by human observers.
2. MSI-Net (Kroner et al., 2019) uses a combi-
nation of encoder-decoder convolutional neural
networks at several levels of granularity. It has
been pre-trained with human eye-tracking data
on a very large corpus of natural images, but
not paintings.
We found that the MSI algorithm works best as an
approximation of what human users find the most
salient area in the Tuymans’ paintings, and this in
turn is often, but not always, a strong cue of the focal
point. The result for the painting ‘Twentyseventeen’
is shown in Fig. 2. It is the face of a woman and the
most salient area is the right eye (from our perspec-
tive), turned towards the viewer. It is clearly the focal
area of the painting. There is a slight secondary focus
on the lips.
Fig.5 shows the application of different algorithms
on ‘The Valley’ shown in Fig.4. MSI-net (left top)
gives the best results focusing on the eyes as primary
salient region, and on the top hairline as secondary
area, emphasizing the unusually big forehead of the
boy. POOLNET shows the whole object and is there-
fore less interesting with respect to identifying the
focal area. SRB shows several regions so that it is
less relevant for finding the most salient one and Fine-
Grained shows edges instead of regions.
Finally, Fig.6 shows the applications of the
salience estimation algorithms for ‘The book’ (Fig.1).
We see that both MSI and Fine-Grained detect a
salient area in the middle of the painting, which is
also concordant with our human experience. It draws
attention to the fold mark (which is confirmed by
comments in the Palazzo Grassi catalog (Donnadieu,
2019)). However, interestingly enough, when the
Perceiving the Focal Point of a Painting with AI: Case Studies on Works of Luc Tuymans
899
Figure 5: Application of different algorithms to detect
saliency for the painting shown in Fig. 4. From left to right
and top to bottom, we used 1. MSI-net. 2. Source image, 3.
SRB, 4. POOLNET, 5. Source-Image, 6. Fine-grained.
Figure 6: Saliency detection using various algorithms on
‘The book’ shown in Fig. 1. From left to right. 1. MSI-net.
2. Fine-grained. 3. SRB. POOLNET gives no results at all
because no object could be detected.
painter Luc Tuymans was shown this result, he re-
marked that there is actually another focal point.
Indeed, when your gaze follows the vertical line
of the fold (see Fig. 1), it ends up at the top edge
of the painting. In that region it becomes suddenly
clearer that these are the pages of an open book, and
that the vertical line is a fold mark. In fact, if one
looks more carefully one sees that the pages on the
top of the vertical line are slightly curled, presumably
to draw further attention to the fact that we are dealing
with an opened book (see Fig.7). This shows the de-
gree of sophistication with which Tuymans attempt to
manipulate the viewer’s gaze and the enormous chal-
lenge to build AI systems that are sensitive to these
art-making strategies, let alone use themselves these
strategies to create new art works.
Figure 7: Tuymans (middle) discusses the importance of the
vertical line in the painting ’The Book’ (Fig. 1), seen here
in the background. The pictures has been taken during con-
versations at the Palazzo Grassi in september 2019 with Luc
Steels (shown to the left) and Massimo Warglien (shown to
the right). This image illustrates the large, almost human-
sized, height of the painting, emphasizing the monumental
character of the church.
4 CONCLUSIONS
This paper has sketched different levels of percep-
tion and interpretation for art works, more specifically
paintings, referring to the earlier influential writings
of Panofsky. Using AI methods, we can unpack these
levels and try to build very precise operational mod-
els in order to shed new light on art, build tutoring
tools for art education, and give a novel instrument to
artists to reflect on their artistic practices, which are
today based on very powerful intuitions and intense
creativity but not on scientific knowledge. The paper
illustrated this approach with the work of the Flemish
painter Luc Tuymans. We conducted a preliminary in-
vestigation on the perception of the focal point, show-
ing the strength and limitations of salience estimation
methods from computer vision and the need to take
other dimensions (such as color usage or the implicit
lines in a painting) as well as semantic issues (such as
triggered by the title or the catalogue) into account.
ACKNOWLEDGEMENTS
This research was made possible by the art-science
initiative of the H2020 FET Proactive project Ody-
cceus with the Ca’Foscari University of Venice as a
ICAART 2020 - 12th International Conference on Agents and Artificial Intelligence
900
partner, the Scientist in Residency project of ‘Luc
Steels in Studio Luc Tuymans’ as part of the EU
H2020 Regional STARTS Center at the Centre for
Fine Arts (BOZAR) in Brussels, and the EU H2020
Humane AI Flagship preparation project of which
IBE (UPF/CSIC) in Barcelona is a partner. This paper
was influenced by a conversation between Luc Steels
and Luc Tuymans at BOZAR in Brussels on 6 novem-
ber 2019. This conversation was a collateral event of
BNAIC, Belgian-Dutch AI conference. The help of
Bram Bots and Isadora Callens of Studio Luc Tuy-
mans in Antwerp for the provisioning of data is grate-
fully acknowledged as well as the help of Christophe
De Jaeghere, Lise Ninane and Emma Dumartheray of
the Gluon Foundation in Brussels for facilitating this
art-science collaboration. The paper was stimulated
by a presentation by LS at the Computer and Art sym-
posium at ECLT, Ca’Foscari University of Venice,
October 2019. LS has been responsible for the re-
search direction and the writing of the paper and BW
has performed technical research into the algorithms
and carried out the experiments. Conversations with
Massimo Warglien from the Ca’Foscari University of
Venice as well as the time and ideas generously pro-
vided by Luc Tuymans are gratefully acknowledged.
REFERENCES
Cetinica, E., Lipica, T., and Grgicb, S. (2018). Fine-tuning
convolutional neural networks for fine art classifica-
tion. Expert systems with applications, 114:107–118.
de Boer, V., Wielemaker, J., van Gent, J., Oosterbroek,
M., Hildebrand, M., Isaac, A., van Ossenbruggen, J.,
and Schreiber, G. (2013). Amsterdam museum linked
open data. Semantic Web, 4(3):237–243.
Donnadieu, M. (2019). La Pelle, exhibition guide. Palazzo
Grassi, Venice.
Garcia, N. and Vogiatzis, G. (2018). How to read paint-
ings: Semantic art understanding with multi-modal re-
trieval. In Proceedings of the 2018 European Confer-
ence on Computer Vision (ECCV). Computer Vision
Foundation.
Girshick, R., Donahue, J., Darrell, T., and Malik, J. (2013).
Rich feature hierarchies for accurate object detection
and semantic segmentation.
Hou, X. and Zhang, L. (2007). Saliency detection: A spec-
tral residual approach. In 2007 IEEE Conference on
Computer Vision and Pattern Recognition. IEEE.
Kroner, A., Senden, M., Driessens, K., and Goebel, R.
(2019). Contextual encoder-decoder network for vi-
sual saliency prediction. CoRR, abs/1902.06634.
Liu, J., Hou, Q., Cheng, M., Feng, J., and Jiang, J. (2019). A
simple pooling-based design for real-time salient ob-
ject detection. CoRR, abs/1904.09569.
Marder, L. (2019). Why focal points in paint-
ings are so important. liveaboutdotcom,
”https://www.liveabout.com/all-about-focal-points-
in-painting-4092634”(2019).
Meyer-Hermann, E. (2019). Catalogue Raisonn
´
e of Paint-
ings. David Zwirner Books, Yale University Press,
New Haven.
Montabone, S. and Soto, A. (2010). Human detection us-
ing a mobile platformand novel features derived from
a visualsaliency mechanism. Image and Vision Com-
puting, 28(3):391–402.
Panofsky, E. (1939,1972). Studies in Iconology. Humanistic
themes in the art of the Renaissance. Oxford Univer-
sity Press, Oxford.
Redmon, J., Divvala, S., Girshick, R., and Farhadi, A.
(2015). You only look once: Unified, real-time ob-
ject detection.
Schmitt, P. (2018). A computer walks into a gallery. In
ICCV Computer Vision Art Gallery. Computer Vision
Foundation.
Semmo, A., Isenberg, T., and Doellner, J. (2017). Neural
style transfer: A paradigm shift for image-based artis-
tic rendering? In Proceedings of the Symposium on
Non-Photorealistic Animation and Rendering. ACM.
Strezoski, G. and Worring, M. (2017). Omniart: Multi-
task deep learning for artistic data analysis. CoRR,
abs/1708.00684.
Tuymans, L. (2018). The Image Revisited. In conversation
with G. Boehm, T. Clark and H. De Wolf. Ludion,
Brussels.
You, Q., Luo, J., Jin, H., and Yang, J. (2015). Robust image
sentiment analysis using progressively trained and do-
main transferred deep networks. In Proceedings of the
29th AAAI Conference. AAAI.
Perceiving the Focal Point of a Painting with AI: Case Studies on Works of Luc Tuymans
901